面向个性化推荐的node2vec-side融合知识表示

node2vec-side fusion knowledge representation for personalized recommendation

推荐系统中知识图谱对系统的推荐效果起到很重要的作用,图谱中的知识表示成为影响推荐系统的关键因素,这也成为当前的研究热点之一。针对推荐系统中知识图谱的结构特点,在传统node2vec模型基础上增加关系表示和多元化游走策略,提出一种基于node2vec的知识表示node2vec-side,结合推荐系统知识图谱网络结构,旨在挖掘大规模推荐实体节点间潜在的关联关系,降低表示方式复杂度,提高可解释性。经过时间复杂度分析可知,提出的知识表示方式在复杂度上低于Trans系列和RGCN。在传统知识图谱数据集FB15K、WN18和推荐领域数据集MovieLens-1M、Book-Crossing、Last.FM上分别进行链接预测对比实验。实验结果表明:在MovieLens-1M数据集上,hits@10分别提升了5.5%~12.1%,MRR提升了0.09~0.24;在Book-Crossing数据集上,hits@10分别提升了3.5%~20.6%,MRR平均提升了0.04~0.24;而在Last.FM数据集上,hits@1提升了0.3%~8.5%,MRR平均提升了0.04~0.16,优于现有算法,验证了所提方法的有效性。

The knowledge graph in the recommendation system plays a vital role in the recommendation effect of the system,and the knowledge representation in the graph becomes a key factor affecting the recommendation system,which has become one of the current research hotspots.This paper proposed a node2vec-based knowledge representation node2vec-side based on the traditional node2vec model by adding relational representation and diversifing wandering strategy to the structural characte-ristics of the knowledge graph in recommendation system,which combined with the knowledge graph network structure of recommendation system to explore the potential association relationship between nodes of large-scale recommendation entities,reduced the complexity of the representation and improved interpretability.After time complexity analysis,it could be seen that the proposed knowledge representation is lower than Trans series and RGCN in terms of complexity.Link prediction experiments were conducted on the traditional knowledge graph datasets FB15K,WN18,and recommendation domain datasets MovieLens-1M,Book-Crossing,Last.FM respectively.The experimental results show that on the MovieLens-1M dataset,hits@10 improves 5.5%~12.1%and MRR improves 0.09~0.24,respectively.On the Book-Crossing dataset,hits@10 improves 3.5%~20.6%,and MRR improves 0.04~0.24 on average,respectively.And on the Last.FM dataset,hits@1 improves 0.3%~8.5%and MRR improves 0.04~0.16 on average.It is better than the existing algorithms and verifies the effectiveness of the proposed method.